This Week in AI: OpenAI’s Data Agent, AI Tackles a 90-Year Maths Problem, Microsoft Copilot Controls and Mistral’s €3bn Bet
This Week in AI: OpenAI’s Data Agent, AI Tackles a 90-Year Maths Problem, Microsoft Copilot Controls and Mistral’s €3bn Bet
The biggest AI news for businesses this week, and what SMB leaders should do with it.
AI is moving closer to real data, real systems, real decisions and increasingly complex problems. For SMBs, that makes permissions, verification and cost control harder to treat as an afterthought. Pick narrow workflows, define what the system can access and act on, and measure the full economics before scaling.
When we looked at what unfolded over the past week, one theme kept coming up: AI is starting to look much more like business infrastructure.
OpenAI pushed its newest model further into serious workplace tasks, launched an agent that can interrogate company data, packaged specialist tools for financial services, and gave developers more of the machinery needed to build long-running and voice-based agents.
Then came something rather different. OpenAI revealed that an internal AI system had produced a proposed solution to one of mathematics’ most famous unsolved problems, a question that has stood for roughly 90 years.
Elsewhere, Microsoft tightened identity and age controls around Copilot, while Mistral raised €3bn to expand its position in sovereign AI.
We’re in for an instalment that looks at vastly different announcements, but they all have a similar direction of travel, in that AI is moving closer to real data, real systems, real decisions and increasingly complex problems. While that makes it more useful as a concept, it also makes permissions, verification and cost control much harder to treat as an afterthought.
Headlines at a glance
- OpenAI’s GPT-6 Astra is rolling into workplace use, with stronger computer use and end-to-end document, spreadsheet and presentation work, but premium capability comes with premium consumption.
- The new Data agent in ChatGPT Work can analyse governed company data, build dashboards and propose actions across connected business tools.
- An internal OpenAI system has produced a proposed solution to the Navier-Stokes Millennium Prize Problem, a mathematical question that’s remained unresolved for roughly 90 years.
- ChatGPT for Financial Services bundles premium market data, financial reasoning and firm templates into a sales-led product for eligible institutions.
- The Agents API and GPT-Live-1 make production agents easier to build, including long-running cloud tasks and natural phone-based conversations.
- Microsoft now requires sign-in for Copilot and restricts access for children under 13, with wider implications for identity and age-aware product design.
- Mistral’s €3bn raise and Cloudera partnership strengthen the case for open-weight, hybrid and sovereign AI deployments.
OpenAI GPT-6 Astra moves from model launch to workplace rollout
OpenAI has positioned GPT-6 Astra as a model for complex, multi-step professional work rather than just packaging it as ‘better answers!’ They say it can use websites and desktop tools, update customer records, conduct research, and produce documents, spreadsheets and presentations that follow supplied templates.
Broadly speaking, it means that the useful unit of AI is changing away from answer quality into if (and how) a system can take a messy assignment, work across several tools, keep track of changing requirements and return something usable.
For departments such as operations, finance, marketing and internal IT teams, it’s going to be something to keep an eye on.
There’s a cost point to watch, however. OpenAI’s API page lists Astra at $10 per million input tokens and $50 per million output tokens, with higher rates for prompts above 272,000 input tokens. ChatGPT Work and Codex allowances also vary by plan and task. A difficult, long-running job can therefore require a much larger allowance than a short drafting request.
Don’t roll a frontier model out as the default for every task. Test it on a small set of jobs where stronger judgment, computer use or finished deliverables could justify the extra cost. Keep approval points around actions in live systems, and compare the full result, including checking and correction time, against a cheaper model or the existing process.
ChatGPT’s Data agent brings governed analytics into the conversation
OpenAI’s new Data agent in ChatGPT Work connects to company data, investigates changes and creates interactive dashboards through plain-language instructions. OpenAI lists connections spanning AWS, ClickHouse, Databricks, Snowflake, MongoDB, Redis and G2, alongside BI tools including Power BI, Tableau, Sigma and ThoughtSpot.
So, not only can someone ask, “Why did sales fall last month?” and get an answer, but administrators also have the ability to choose which data connections and roles are available, while queries retain the connected account’s table, row and column permissions. The agent can also recommend next steps and send findings through Slack or email, with actions subject to approval.
That puts analytics within reach of more people, particularly smaller teams that don’t have a dedicated analyst waiting for every question… But be wary that it also exposes weak foundations very quickly.
If the business has inconsistent metric definitions, excessive permissions or unreliable source data, a polished dashboard won’t repair them; it may simply make the problem look more convincing.
Begin with one well-understood dataset and a question whose answer can be checked independently. Define the metric, confirm who should see the underlying records, and compare the result with an existing report before letting the workflow trigger emails or actions. The agent can accelerate governed data work, but it can’t replace the underlying governance.
An OpenAI system takes on one of mathematics’ biggest unsolved problems
This week’s most eye-catching AI story didn’t involve another chatbot feature, but rather an impressive demonstration of AI’s capabilities.
On the 8th of September, OpenAI published a proposed solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems established by the Clay Mathematics Institute in 2000.
The underlying question has remained unresolved for roughly 90 years. Now, maths classes were almost alien to me for a large part of my time in school, so bear with me as I try to explain what this is, if you’re unfamiliar.
As I understand it,the Navier-Stokes Problem concerns the equations used to describe how fluids move and whether smooth three-dimensional fluid motion can eventually develop a singularity, where the mathematical behaviour effectively breaks down. (Thank you to ChatGPT for the assist on that one)
Make sense? No, it doesn’t to me either.
I digress. Let’s keep it moving.
OpenAI says its system produced an analytical proof showing that a singularity can develop in finite time, alongside a formalisation in Lean that allows the mathematical proof to be checked computationally.
But here’s the actual most interesting part: what produced it.
OpenAI says it used a system of coordinating agents powered by an internal model that’s still being trained and is “significantly more capable” than GPT-6 Astra. Around 10,000 concurrent agents were involved in the group that produced the Navier-Stokes result, with the agents arriving at their solution after roughly 88 hours. GPT-6 Astra was then used as part of the Lean formalisation and verification process.
OpenAI isn’t claiming the $1m Millennium Prize, and the work will face scrutiny from the mathematical community. But the significance goes beyond whether most businesses will ever need an AI system to solve advanced fluid dynamics.
If nothing else, it’s further evidence of where frontier AI development may be heading. Systems are increasingly being designed to investigate difficult problems over long periods, coordinate large numbers of agents, use specialist tools and produce work that can be independently checked.
You probably don’t have a 90-year-old maths problem sitting in your backlog, so broadly speaking, you may think this is a bit of a novel footnote - or even question why I’ve stuck this in here. But honestly, there is actually a useful lesson here in the approach. AI is moving from answering isolated questions towards working through much larger problems using multiple steps, tools and verification. For businesses, the ability to define a problem clearly, give AI the right information, and independently check the result is increasingly important.
ChatGPT for Financial Services packages data, models and templates
OpenAI also launched ChatGPT for Financial Services, a tailored ChatGPT Work experience shaped with Morgan Stanley and Evercore. It combines GPT-6 Astra with built-in datasets from providers such as Daloopa, PitchBook and LSEG News, and supports workflows including company research, financial modelling, earnings analysis and PitchBook preparation.
There’s a sensible product idea underneath the sector label, as the hard part of specialist AI often isn’t the model - it’s getting reliable data, entitlements, templates, permissions and evidence into one workflow.
OpenAI says users can trace claims to supporting tables and passages, while administrators can publish approved Excel, Word and PowerPoint templates for their teams.
The governance features include SAML single sign-on, SCIM provisioning, role-based access controls, configurable retention and supported log exports. For those concerned, OpenAI says business data isn’t used to train its models by default.
Availability is currently limited to eligible financial institutions via sales or account teams, so this isn’t an off-the-shelf add-on for every finance department.
There’s another caveat worth making explicit: OpenAI’s financial-services terms say partner data and outputs may be inaccurate, delayed or incomplete, and don’t replace professional judgment.
Accountants, advisers, fintechs and regulated firms should watch the packaging rather than rush to buy. AI needs licensed data, clear rights, auditable sources, approved templates and human accountability, but a general chatbot with a finance prompt isn’t the same thing as a fleshed out department.
OpenAI’s Agents API and GPT-Live-1 lower the build barrier
For developers and software businesses, the Agents API may be the more consequential launch.
In public beta, it provides a managed harness for agents that need to retain context, call tools, coordinate subagents and continue working across long sessions. Developers can use an OpenAI-hosted sandbox, their own infrastructure or supported environment partners. OpenAI says there’s no separate API fee, as customers pay for the tokens and tools used.
OpenAI also released GPT-Live-1 in the API for full-duplex voice agents that can listen and speak at the same time, handle interruptions and delegate deeper reasoning to a back-end model. It supports telephony and is priced at $0.05 per minute for the front-end voice layer, before the cost of the reasoning model and agent tools behind it.
Taken together, these products make it easier to build agents that investigate, act and converse, which could be valuable for customer service, booking, field support, internal IT and workflow automation.
But it does make architecture and cost slightly more ambiguous. A voice minute may trigger model tokens, search, data retrieval, tool calls and a longer-running task behind the scenes, so it’s something to be mindful of.
Start with a tightly bound workflow where the system can hand over cleanly to a person. Log the tools it calls, cap what it can spend or change, and test interruption, failure and escalation paths. The demo shouldn’t be the acceptance test, but an awkward real customer (or someone posing as one) should be.
Microsoft tightens Copilot identity and age controls
Microsoft’s new Safe Participation Framework says all Copilot users must now sign in, and access is restricted for children under 13 (or maybe even older, where local law requires it).
Microsoft also points to a Windows Age API intended to provide privacy-conscious age signals to developers building age-appropriate experiences.
For most office-based SMBs, this isn’t a Copilot productivity upgrade, but it’s still commercially relevant because it shows identity becoming a prerequisite for safer AI use.
For schools, training providers, charities, family-facing services and software companies, age assurance and differentiated experiences are moving from policy discussion into product design.
If staff or customers may include children or teenagers, check how AI access is authenticated, what age signals are used, which data is collected and what happens when age can’t be established. Even for adult-only organisations, the direction is still useful as anonymous access and meaningful governance rarely coexist comfortably.
Mistral’s €3bn raise strengthens the sovereign AI alternative
European AI company Mistral announced a €3 billion Series D at a post-money valuation above €21 billion, led by Samsung Electronics with Scaleup Europe Fund and PSG Equity as co-leads.
Mistral says the funding will expand research, computing, infrastructure and international growth. It currently operates across 20 countries and supports more than 125 global enterprises.
Two days later, Mistral and Cloudera announced a partnership for sovereign enterprise AI. The planned integration is designed to run inference across private cloud, public cloud, on-premises and fully air-gapped environments, while allowing organisations to customise models against proprietary data inside controlled environments.
This isn’t an immediate buying signal for the average small business, but we view it as a sign that the enterprise AI market isn’t settling into a simple OpenAI-versus-Anthropic choice.
Data location, open weights, deployment control, auditability and vendor dependence are becoming competitive features in their own right, particularly in regulated sectors and Europe.
Add portability and data control to AI procurement questions now, even if you’re buying a managed service. Ask how data can be exported, whether workflows depend on proprietary connectors, where inference runs and what happens if pricing or availability changes. Sovereignty sounds like a government-sized concern until vendor lock-in lands on an SME budget.
Quick answer: What should SMEs do about AI this week?
Choose one data-backed workflow to test: Define the source data, permissions, expected output and a human check before switching anything on.
Separate model access from action authority: A system may be allowed to analyse records without being allowed to email customers, update the CRM or commit spend.
Measure the total economics: Include licences, tokens, tool calls, voice minutes, administration, checking, corrections and escalation time in the business case.
Fifosys view
Perhaps the most prominent story this week might well be an AI system taking on a problem that’s challenged mathematicians for decades (if you can make it make sense to you, that is). For businesses, though, the more immediate story is on the stack forming around the same technology.
Models are being connected to company databases, specialist datasets, office templates, phone lines and production systems. At the frontier, they’re also being combined into large groups of agents capable of working on a single difficult problem for days at a time.
That’s how AI becomes genuinely useful, and how a vague experiment can become an operational dependency.
For growing businesses, we’ll keep repeating this point: don’t chase every launch or wait for the market to become simple. Why? Well, that’s simple. It probably won’t ever be simple.
Pick narrow workflows, give them clean data and clear boundaries, keep a person accountable for the outcome, and track what the work really costs.
The businesses getting the most from AI won’t necessarily have the most advanced model switched on everywhere, but they’ll be the ones who know which model is doing what, which data it can reach, which actions it can take and whether the result is actually better than the process it replaced.
That’s less headline-grabbing than AI solving a famous maths problem, but it’s where the practical value sits.
This week's AI news FAQs
What is OpenAI’s Data agent?
The Data agent in ChatGPT Work is designed to connect to governed company data, investigate questions and create interactive dashboards through natural-language instructions while retaining the permissions of the connected account.
Did AI solve the Navier-Stokes Millennium Prize Problem?
OpenAI has published a proposed solution produced by an internal multi-agent system, alongside a Lean formalisation. The work still needs scrutiny from the mathematical community, and OpenAI isn’t claiming the $1m Millennium Prize at this stage.
What is ChatGPT for Financial Services?
It is a tailored ChatGPT Work experience for eligible financial institutions that combines OpenAI models with specialist financial datasets, firm templates and enterprise governance controls.
What are OpenAI’s Agents API and GPT-Live-1 for?
The Agents API provides infrastructure for long-running agents that retain context and use tools, while GPT-Live-1 is designed for real-time voice agents that can listen, speak, handle interruptions and connect to deeper reasoning workflows.
Why does Mistral’s funding matter to businesses?
Mistral’s €3bn raise and sovereign AI partnership with Cloudera reinforce alternatives to closed, single-provider AI stacks, particularly for organisations that care about data location, deployment control, open weights and vendor dependence.
Build AI around real work, not every new launch
We can help you identify useful workflows, understand the data and permissions behind them, and put practical governance around AI adoption.
Talk to our team
Discuss where AI could create practical value in your organisation and what needs to be in place first.
AI for business
See how Fifosys helps organisations adopt AI around real workflows, clear controls and measurable value.